Epistemic Dimension (v1.1.0)

MIN v1.1.0 introduces Epistemicum as the 6th Forma subclass.

Motivation

Engineering knowledge graphs need to express: - "Engineer Mueller believes Hooke's law holds for DC04 up to 200 MPa." - "This hypothesis is supported by test data with 0.95 statistical confidence." - "Two experts disagree on whether the model is valid."

Before v1.1.0, MIN could express that a process confirms a law (confirms) or refutes a possibility (refutes), but could not attach epistemic status, confidence, or agent attribution to these judgements.

Design decision: why Forma?

Epistemicum is not causally efficacious (it does not transform matter or produce data) — so it is not Nexus. It formally determines how an agent interprets a state of affairs — so it is Forma.

Checked against all existing Forma subclasses:

Subclass Why Epistemicum is different
Lex Lex holds universally. Epistemicum is an agent's stance, not a law.
Structura Structura formalizes. Epistemicum evaluates.
Possibile Possibile is a possibility. Epistemicum is a stance about a possibility.
Norma Norma prescribes what should hold. Epistemicum describes what is held to be true.
Institutio Institutio exists through collective recognition. Epistemicum can be held by a single agent.

Core mechanics

Epistemicum has five building blocks that combine freely:

  • WHO believes it? → heldBy Agent (or agent-free: institutionalized)
  • WHAT is believed? → about Entity (Lex, Norma, Data, Object, Process — anything)
  • HOW CERTAIN? → hasConfidence [0..1] + hasConfidenceType (Subjective / Statistical / Bayesian / Heuristic)
  • WHICH STATUS? → hasEpistemicStatus (Hypothetical / Confirmed / Refuted / Contested / Axiomatic)
  • BASED ON WHAT? → supportedBy / underminedBy Nexus (measurements, experiments, calculations)

All patterns below arise from these five building blocks.

Two complementary epistemic patterns

MIN v1.1.0 maintains two distinct epistemic patterns:

Pattern 1: Process-centric (Popperian, v1.0.0)

ex:TensileTest  min:confirms  ex:HookesLaw .
ex:TensileTest  min:refutes   ex:LinearityAbove300MPa .

The process is the carrier. Semantics: a process evaluates a formal determinant as tenable or untenable.

Pattern 2: Evidence-centric (Epistemicum, v1.1.0)

ex:Hypothesis  a min:Epistemicum ;
    min:heldBy       ex:Mueller ;
    min:about        ex:HookesLaw ;
    min:supportedBy  ex:TestData ;
    min:hasEpistemicStatus  min:ES_Confirmed ;
    min:hasConfidence       0.95 ;
    min:hasConfidenceType   min:CT_Statistical .

The epistemic stance is the carrier. Semantics: a reified belief accumulates evidence, has status and confidence.

Both patterns are needed in practice. Pattern 1 is compact and sufficient for simple process-level assertions. Pattern 2 is needed when epistemic state changes over time, when confidence must be typed, or when multiple agents hold conflicting stances.

EpistemicStatus values

Value Meaning
ES_Hypothetical Proposed, neither confirmed nor refuted
ES_Confirmed Supported by evidence and accepted
ES_Refuted Contradicted by evidence
ES_Contested Conflicting evidence, no consensus
ES_Axiomatic Foundational assumption, not empirically testable

ConfidenceType values

Value Meaning
CT_Subjective Expert judgement
CT_Statistical From sample (confidence interval)
CT_Bayesian Posterior probability
CT_Heuristic Experience-based estimate

Engineering patterns

Model validation. An FEM engineer models: "I believe the Euler-Bernoulli beam theory (Structura) holds for this load case. Confidence 0.7 (HeuristicConfidence). Supported by simulation data. Undermined by measurement data at high load." This is V&V as a graph instead of a report.

Material release. A test engineer models: "The tensile test (Process) confirms that Rm >= 270 MPa (Norma) is met. Status: Confirmed. Confidence 0.98 (StatisticalConfidence). Supported by 30 measurements." The acceptance criterion becomes queryable: "Which Norma instances on this part are Confirmed?"

Failure analysis. An expert models: "Hypothesis: Corrosion (Process) caused the failure. Status: Contested. Agent A says Confirmed (supported by metallography), Agent B says Refuted (supported by fracture surface analysis)." Dissent becomes visible, not hidden.

Simulation chain trust. In a process chain (deep drawing -> heat treatment -> assembly) each step has its own uncertainties. Epistemicum allows: one hypothesis with confidence per simulation step. The final question: "Which step has the lowest confidence?" — validate there first.

Material model selection. "For PP-GF30 I use the Mori-Tanaka model (Structura). Status: Hypothetical. Confidence 0.6 (SubjectiveConfidence). Alternative: RVE homogenization (second Epistemicum, about the same Structura, higher confidence)." Decisions become reasoned and comparable.

DPP patterns (Digital Product Passport)

Auditable recyclate content. Not "42%", but: "42%, Epistemicum: Confirmed, Confidence 0.92 (StatisticalConfidence), supportedBy: XRF measurement data, heldBy: test lab TUeV." An auditor can ask: "Which DPP entries have confidence < 0.5?" or "Which are only SubjectiveConfidence?"

SVHC declaration. "This product contains no SVHC above 0.1%. Status: Confirmed. supportedBy: supplier declaration (Data)." Versus: "Status: Hypothetical. supportedBy: none." The difference is regulatorily relevant.

Supply chain trust. Each actor in the chain adds Epistemicum instances. Raw material supplier: "Origin: Virgin. Confidence 0.99." Recycler: "Origin: Recycled, 60% share. Confidence 0.85." OEM: can query the entire chain and identify the weakest link.

DPP completeness analysis. Query: "Which mandatory fields in the DPP have no associated Epistemicum?" — the gaps become visible. "Which have Status Assumed instead of Confirmed?" — the weak spots too.

Digital twin patterns

Sensor credibility. A sensor (HardwareAgent) delivers temperature data (Data). Epistemicum: "This data is correct. Status: Confirmed. Confidence 0.95." After calibration drift: "Status: Contested. Confidence 0.4. underminedBy: reference measurement." The digital twin knows which sensors to trust.

Predictive maintenance. An ML model (SoftwareAgent) says: "Component fails within 200 operating hours (Possibile). Confidence 0.72 (BayesianConfidence). supportedBy: vibration data from the last 30 days." The maintenance decision has a formal basis.

Model update. Old hypothesis: "Linear degradation model holds." New hypothesis supersedes old, with higher confidence, supported by more data. The history is queryable via property paths: "What did we believe 6 months ago, and why not anymore?"

Scientific patterns

Hypothesis -> experiment -> result. The classic scientific cycle becomes a graph pattern:

  1. Epistemicum (Hypothetical, about Lex)
  2. Process (Experiment) generates Data
  3. Data supportedBy/underminedBy Epistemicum
  4. New Epistemicum (Confirmed/Refuted) supersedes old

Peer review. Reviewer A: Epistemicum about paper thesis, Confirmed. Reviewer B: Epistemicum about same thesis, Refuted. Paper status: Contested. Queryable: "Which theses in my research project are contested?"

Reproducibility. Original study: Epistemicum, Confirmed, supportedBy Experiment_1. Replication study: new Epistemicum, underminedBy Experiment_2. Replication status is formalized.

Meta-analysis. Multiple Epistemicum instances about the same Lex, different supportedBy datasets, different confidences. The meta-question "How well is F=ma supported for this parameter range?" becomes an aggregation query.

Organizational patterns

Knowledge management. "Department A believes procedure X is optimal. Department B prefers Y." Both formalized as Epistemicum with different supportedBy evidence. Management can ask: "Where is there dissent between departments?"

Lessons learned. A project error modeled as Epistemicum: "We assumed (Axiomatic) that supplier X delivers on time. Status: Refuted. underminedBy: delay data." The lesson learned is queryable, not buried in a PDF.

Onboarding. A new employee can ask: "What is Confirmed in our knowledge graph? What is still Hypothetical?" Instead of implicit knowledge, the epistemic state of the team becomes visible.

Standards patterns

Standard evolution. DIN EN ISO 6892-1:2019 supersedes DIN EN ISO 6892-1:2009. Both as Institutio. The community's epistemic stance: old version Refuted (no longer valid), new version Confirmed. Queryable: "Which standards in my system are deprecated?"

Compliance gaps. Epistemicum: "We fulfill Norma X. Status: Confirmed." Versus: "We fulfill Norma Y. Status: Hypothetical, Confidence 0.3, supportedBy: none." The difference between "tested" and "claimed" becomes explicit.

Knightian / deep uncertainty

Known unknowns. isQuantifiable = true on Possibile: "Component failure with 2% probability." Normal risk, quantifiable.

Unknown unknowns. isQuantifiable = false on Possibile: "Climate change impact on supply chain." No distribution assignable. This is not a bug but a deliberate statement: "We don't know what we don't know — but we know that we don't know it."

Scenario planning. Three Possibile instances (best case, base case, worst case), each with its own Epistemicum and different confidence. Queryable: "Which scenario has the most evidence?"

Agent optionality

heldBy has no minimum cardinality in SHACL. Agent-free Epistemicum instances model institutionalized epistemic states:

ex:DatasetVerified  a min:Epistemicum ;
    min:about  ex:Dataset_042 ;
    min:hasEpistemicStatus  min:ES_Confirmed .
    # No heldBy — institutional state, not personal belief.

Provenance for agent-free instances runs through the existing originatedBy.

Graph patterns (SPARQL)

# Lowest confidence in the system — weakness analysis
SELECT ?e ?subject ?confidence WHERE {
    ?e a min:Epistemicum ; min:about ?subject ; min:hasConfidence ?confidence .
} ORDER BY ASC(?confidence) LIMIT 10

# Dissent between agents
SELECT ?forma ?a1 ?a2 WHERE {
    ?e1 min:about ?forma ; min:heldBy ?a1 ; min:hasEpistemicStatus min:ES_Confirmed .
    ?e2 min:about ?forma ; min:heldBy ?a2 ; min:hasEpistemicStatus min:ES_Refuted .
    FILTER (?a1 != ?a2)
}

# Knowledge progress: what was confirmed recently?
SELECT ?e ?forma WHERE {
    ?e min:about ?forma ; min:hasEpistemicStatus min:ES_Confirmed .
    ?e_old min:about ?forma ; min:hasEpistemicStatus min:ES_Hypothetical .
    ?e min:supersedes ?e_old .
}

# Key experiments: which evidence supports the most hypotheses?
SELECT ?evidence (COUNT(?e) AS ?count) WHERE {
    ?e min:supportedBy ?evidence .
} GROUP BY ?evidence ORDER BY DESC(?count)

# Blind spots: non-quantifiable Possibile
SELECT ?p WHERE {
    ?p a min:Possibile ; min:isQuantifiable false .
}

# Unsupported assumptions: Epistemicum without evidence
SELECT ?e ?subject WHERE {
    ?e a min:Epistemicum ; min:about ?subject .
    FILTER NOT EXISTS { ?e min:supportedBy ?x }
    FILTER NOT EXISTS { ?e min:underminedBy ?x }
}

Design decisions

Why no EM_Epistemic

EfficacyMode tracks the mode of determination — how a Forma subclass constrains Nexus. Each EfficacyMode individual maps 1:1 to a Forma subclass (lawful = Lex, structural = Structura, etc.).

Epistemicum does not determine — it evaluates. Adding EM_Epistemic would conflate two orthogonal axes. The epistemic dimension is correctly captured by hasEpistemicStatus and hasConfidenceType.

Why supersedes stays non-transitive

supersedes was explicitly declared non-transitive in v1.0.0. The rationale remains valid: A supersedes B and B supersedes C does not imply A directly supersedes C (intermediate versions may change context).

For history queries, use SPARQL property paths:

SELECT ?latest ?earliest WHERE {
    ?latest min:supersedes+ ?earliest .
}

SHACL constraints

Constraint Severity
about min 1 Violation
hasEpistemicStatus exactly 1 Violation
hasConfidence in [0.0, 1.0] Violation
hasConfidence present implies hasConfidenceType present Violation
supportedBy / underminedBy must point to Nexus Violation
heldBy must point to Agent (if present) Violation

What Epistemicum CANNOT do

No truth. Epistemicum says "Agent A believes X with confidence 0.9." It does not say "X is true." Truth is not a MIN concept — and that is the right call.

No reasoning over confidence. A reasoner cannot infer: "If confidence(A) = 0.8 and confidence(B) = 0.7, then confidence(A and B) = 0.56." Bayesian propagation is a Process, not an axiom.

No emotions. "I feel that the component will fail" is not an Epistemicum. Epistemicum is propositional: it refers to states of affairs, not feelings.

No self-reference. "I believe that I believe" is not cleanly modelable in OWL. Epistemicum can talk about anything in the graph — except itself.